The architecture, in seven claims
Your IP never trains public models
Inference-only contracts with every upstream model vendor. Zero training retention — your scripts, frames, and sequences are never used to train anyone's models.
Isolation enforced at the database layer
Every table is fenced by row-level security — 361 policies across 84 tables — and assets live under tenant- and project-scoped paths. Isolation is a database guarantee rather than an application promise. Where a bucket is public, those storage policies govern authenticated access rather than the public read path; private delivery is a dedicated-deployment property.
Encrypted in transit and at rest
TLS 1.3 in transit, AES-256 at rest under the managed platform's key service. Customer-managed keys are a funded roadmap item for dedicated and VPC deployments, not a control we hold today.
Right to delete, and it reaches the files
A deletion request removes your database records and every stored object under your tenant's paths — frames, videos, uploads, models — and the run reports the object count it erased. Backups age out of the recovery window within 30 days.
Vendor-neutral by architecture
1,400+ engines behind one orchestration layer. A production can be limited to the engine makers and jurisdictions its contract allows: tell us what your legal team excludes and we set it on your productions.
Auditable by design
Append-only audit logging on administrative actions, per-generation records, and full asset provenance — exportable for your compliance review.
Continuity that's built, not promised
Automated nightly backups replicate the database, credential-preserving auth records, and every customer asset to independently controlled infrastructure in a separate provider account, with a standby deployment of the application proven on every release. Recovery objectives are documented — a 15-minute recovery point, a 4-hour database restore — and the first full restore rehearsal is scheduled; the log is part of the diligence pack.
How your work stays out of training sets.
"Zero training retention" is not a slogan here — it is a set of enforcement points in the architecture. Your scripts, frames, and sequences are never used to train our models or anyone else's. Here is what actually prevents it.
Commercial API inference only
Every generation call goes to model vendors' commercial API endpoints under terms that exclude the use of inputs and outputs for model training. No consumer apps, no free tiers, no data-for-discount programs — the tiers where training-use terms live are never in the pipeline.
One chokepoint, server-side keys
Your browser never talks to a model vendor. Every call is brokered by our orchestration layer with credentials that exist only server-side — so the complete list of vendors that can ever see a frame is known, versioned, and enforceable. There is no shadow path for content to leak into a training pipeline we didn't contract with.
Process-and-return, not store-and-learn
Vendors process a prompt or frame solely to return your output; your work persists only in your tenant-scoped storage on our side. Where a vendor retains short-lived logs for abuse prevention, the same no-training terms apply to those logs.
Named-vendor exclusions, enforced by the router
If your legal team excludes a provider — any provider — the router guarantees no request from your tenant ever reaches it. Exclusion is a routing rule, not a policy memo.
Provenance you can hand to your client
Lineage records which engine produced every asset, automatically. You can show exactly which vendors processed which frames on any project — the evidence trail a rights holder or compliance review asks for.
And if the answer must be 'no third party at all'
VPC and air-gapped designs run open-weight models inside your own perimeter, with nothing leaving it — the orchestration layer works the same and the vendor list is simply empty. Both are on the roadmap rather than provisioned today, scoped and stood up with our engineering team per engagement.
What we don't do, in plain terms: we do not train models on customer content, we do not fine-tune shared models with customer data, we do not sell or share customer content, and we never route generation through consumer-tier endpoints. These controls map to the SOC 2 Confidentiality criteria; the vendor register and contract terms are published with the policy pack for any security review.
Frames made in Previz. Material at this stage is unreleased, which is what the isolation and retention rules above exist to protect.
From shared cloud to air-gapped.
Shared SaaS
Running todayMulti-tenant. Path-isolated storage, row-level security, shared inference pool. Fastest onboarding.
Dedicated Tenant
In buildSeparate database, storage, and edge functions. Tenant-specific approved engine pool. Named-vendor exclusions enforced. Scoped and stood up per engagement — not a switch we flip.
VPC / Single-Region
Roadmap · fundedDeployed inside a customer-controlled cloud account or region, with customer-managed keys and a network egress allowlist — all three on the funded roadmap rather than built today. No date is quoted until it is scoped with you.
Air-Gapped / On-Prem
ExploringOpen-weight inference on customer hardware. No external network. Full orchestration, no upstream calls.
Security review, without the runaround.
A complete, versioned security documentation pack is published and readable now — no NDA, no waiting on paperwork: system architecture & data-flow diagrams (including every GenAI entry and exit point), the subprocessor register, and our full policy set — access control, data protection & deletion, patch and vulnerability management, secure development, incident response (24-hour customer notification commitment), security awareness training, vendor management, business continuity & disaster recovery, and logging & monitoring. We answer security questionnaires directly, with engineers in the room.
This page is our own statement of practice — an honest self-assessment, not a contract. The legal centre carries a plain-language summary of it. The binding documents are the Terms, Privacy Policy, and Acceptable Use.
Terms of service
The agreement covering accounts, plans, ownership and termination.
/privacyPrivacy policy
What is collected, why it is held, and for how long.
/acceptable-useAcceptable use
The content and the conduct the platform will not carry.
/cookiesCookie policy
Which cookies are set, by whom, and how to refuse the optional ones.
/trustTrust centre
Partner principles, rights and consent, and an open self-assessment.
Security
The architecture: isolation, retention, and deployment modes.
Security documentation
The published policy, architecture and evidence set.
/legalLegal centre
Plain-language summaries of five of these documents, in one room.